Charm++ parallel programming system

[Eric Bohm <ebohm AT illinois.edu>]

There are a wide variety of ways which Charm++ applications might make 
use of threads.  The current TAU interface is somewhat limited in how 
well it identifies them.

The most common case is in an smp based build where pthreads are 
launched at startup and pinned 1:1 to cores (or SMTs) as governed by the 
use of command line arguments.  They then act as the computational 
environment for the Charm++ message driven paradigm without exposing 
pthread semantics to the user.

A slightly less common case uses Charm++ lightweight threads (not 
pthreads) at a ratio higher than the number of hardware execution 
units.   AMPI is the typical venue for this, where each lightweight 
thread holds an MPI rank.   Again, those threads are typically all 
launched at startup rather than being dynamically created/destroyed. The 
lightweight thread scheme can also be used directly via the [threaded] 
semantics of charm++,  an example of this can be found in the ChaNGa 
application.   In those cases, the Projections tool 
http://charm.cs.illinois.edu/manuals/html/projections/manual.html is 
generally useful for performance analysis.

It is also possible to mix OpenMP instructions with Charm++. However, in 
that case the performance of those threads is controlled by the OMP 
infrastructure rather than Charm++.

I have also cc'd Scott Biersdorff on this because I believe he is the 
most familiar with the Charm TAU interface.

On 02/28/2013 12:07 PM, Noel Keen wrote:
> Hello,
>
> I'm trying to get some perf data on a charm++ app.  Looking for the number
> of threads created and the expense of the creation/deletion, etc.  I'm
> trying to build charm++ with TAU on a Cray XT with limited success.
> But if someone knows anything about getting this type of perf data,
> let me know.  Thanks
>
> Noel
> _______________________________________________
> charm mailing list
> charm AT cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/charm